A new economy.
A place for
the small guy.

# AI Memory: From Re-explaining to Working Continuity You return to a promising project, open a fresh conversation, and spend the first ten minutes explaining what you had already decided. The files still exist. The continuity has gone missing. E124 asks how AI can help recover that continuity without quietly taking over the meaning of the work. Zsolt's account begins with the limits he encountered in native assistant memory, moves to external Notion records, and then to agents that can retrieve and connect those records. He reports an archive of roughly 2,000 Suno songs. That scale and the practical benefits are his testimony; this edition has not inspected or independently counted the archive. ## Three different jobs The informational / operational / intentional framework is an E124 synthesis: - **Informational:** what is recorded, where it came from, and whether it can be retrieved. - **Operational:** what is currently approved, who decided it, and which earlier decision it supersedes. - **Intentional:** why the work matters, which constraints protect its purpose, and when a change needs human judgment. A retrieved transcript can help with the first job while failing the other two. A suggestion made last month may resemble today's question but still be rejected, obsolete or outside the user's authority. The account of Jeff Borneman's concern supplied for this production is about authority and shared meaning, not simply insufficient storage. His original post was not supplied, so this edition does not attribute a verified quotation or this three-part framework to him. ## What the research supports LongMemEval separates long-term conversational memory into several testable abilities, including updates and abstention [R1]. Lost in the Middle found that information placement affected performance in its tested long-context tasks [R2]. Neither establishes that every current model fails identically; both motivate testing the actual retrieval job rather than counting stored tokens. Retrieval-augmented generation connects a generator to an external information store [R3]. MemGPT explores managed movement between memory tiers [R4]. These are useful engineering approaches to making past material available. They do not, by themselves, certify that a record is true, current, permitted or approved. OpenAI's February 2024 memory announcement is a dated example of native assistant memory and user controls [R6]. It is not evidence for a universal early 6,000-word capacity, nor a description of every current account. Living features and access conditions change. ## A practical continuity record Begin with a small approved state: the goal, current decision, rationale, owner, timestamp, source links, constraints, unresolved questions and superseded records. Keep proposals visibly separate from approvals. Let an agent retrieve evidence and propose changes; require the responsible person to approve consequential changes. This is an editorial design recommendation, not a validated universal architecture. A small project may only need a short document. A larger archive may benefit from search, metadata and explicit relationships. The relevant question is whether the next session can resume correctly with less total effort. ## Retrieval has costs and boundaries An agent can retrieve the wrong version, turn a draft into an apparent decision, lose the rationale, or repeat a poisoned record. AgentPoison demonstrates that memory and retrieval stores can become an attack surface [R5]. Treat retrieved material as evidence to assess, not new instructions to obey. Limit access to the task, preserve provenance, and make correction and deletion possible. More continuity can also mean more anchoring. A deliberate fresh start may help some creative tasks. The human-memory comparison in E124 is an analogy for selective recall and reinterpretation; it does not establish an evolutionary purpose for forgetting or show that a machine has human intentions. ## Count the whole cost Net human minutes saved = baseline effort avoided − review − maintenance and administration. The source M-report's illustrative numbers require a correction: 250 − 150 − 120 = −20 minutes, meaning **20 extra minutes**, not 130 saved. This is arithmetic on hypothetical inputs, not a measured result. METR's early-2025 developer study is a useful warning about confusing perceived benefit with measured completion time: its selected developers took longer with AI [R7]. It did not test this E124 workflow and cannot determine its value. Measure your own task quality, errors, correction effort and elapsed human time. This edition makes no measured 50:1 claim. ## Five experiments still to run 1. Compare native context with external retrieval using the same answer keys and model. 2. Compare a conversation summary with a concise approved-state record. 3. Compare ordinary briefs with explicit goal, rationale and departure checkpoints. 4. Test archive retrieval on a consented sample before trying creative recombination. 5. Compare continuity-rich and deliberately fresh-start conditions. All five are proposed protocols, not completed experiments. The companion pack supplies templates, scoring fields and stopping conditions. No private archive, audio or full research paper is included. ## What is available The public edition contains the owner-approved 9:28 portrait explainer and this research companion. No Shorts are available for this release. The optional €2.50 research pack adds structured evidence, a 108-source inherited bibliography index, a 133-entry source mapping, five proposed protocols, a decision-state template and a net-time worksheet. Payment is a human checkout handoff through AED. Machine-readable discovery does not support autonomous agent payment. The 108-source Manus bibliography and 133 evidence entries are different units. Only the seven selected primary sources listed below were checked for this edition, at abstract or official-summary depth. The wider index is a discovery aid, not a claim that every source or claim has been independently verified. Full source reports remain with their publishers. ## References - [R1] LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. https://arxiv.org/abs/2410.10813 Review scope: Abstract, version 2. Five tested abilities include extraction, multi-session reasoning, temporal reasoning, knowledge updates and abstention. A benchmark is not a guarantee about today's products. - [R2] Lost in the Middle: How Language Models Use Long Contexts. https://aclanthology.org/2024.tacl-1.9/ Review scope: Published abstract. In the tested retrieval tasks, relevant information position affected performance. More available context does not automatically mean reliable use. - [R3] Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. https://arxiv.org/abs/2005.11401 Review scope: Abstract. The architecture combines a parametric generator with a non-parametric retriever. Retrieval does not establish permission or decision authority. - [R4] MemGPT: Towards LLMs as Operating Systems. https://arxiv.org/abs/2310.08560 Review scope: Abstract, version 2. Hierarchical memory management moves information across tiers to extend usable context. This is a system design, not human memory. - [R5] AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases. https://arxiv.org/abs/2407.12784 Review scope: Abstract. Experiments demonstrate attacks through poisoned memory or retrieval stores. Their rates should not be generalized to every deployment. - [R6] Memory and new controls for ChatGPT. https://openai.com/index/memory-and-new-controls-for-chatgpt/ Review scope: Official announcement and dated updates. The February 2024 announcement describes a limited memory test and user controls. Later updates must not be backdated. - [R7] Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ Review scope: Primary study summary, July 2025. A randomized study of 16 developers and 246 issues found 19% longer completion with AI in that setting. This is not a memory-system experiment or a verdict on current tools. End of E124 public research companion. Edition 1.0, 20 October 2026.